What is optimize based on data?

Optimize based on data is a concept in personal epistemology: Use monitoring data to make targeted improvements to your agents.

Example: You built a morning routine agent six months ago. It fires at 6:00 AM and runs a sequence: meditate, journal, exercise, review today's priorities. The agent worked beautifully for the first two months. Then you noticed — through your monitoring data from Phase 28 — that the agent's reliability had dropped from 92% to 61%. You did not panic. You did not scrap the routine. You looked at the data. The monitoring logs showed a clear pattern: the failure point was almost always the exercise step. Meditation and journaling fired consistently. But exercise failed on Mondays, Wednesdays, and Fridays — the days you had early meetings. The data told you exactly what was broken and when it broke. So you made one targeted change: you moved exercise to the evening on early-meeting days. The next month, reliability climbed back to 88%. A month after that, you noticed the journaling step was producing diminishing returns — your entries were getting shorter and more formulaic. The data showed average journaling time had dropped from twelve minutes to three. So you changed the journaling prompt from a free-write to a structured three-question format. Journaling time stabilized at seven minutes, and the entries became useful again. Two adjustments, both driven by data, both targeted at the specific failure the data revealed. That is optimization: not rebuilding the whole system, but iterating on the parts the evidence says need attention.

This concept is part of Phase 29 (Agent Optimization) in the How to Think curriculum, which builds the epistemic infrastructure for agent optimization.